Personalizing the Search for Persons: A Recommender-
based Approach
Tobias Keim
1
, Jochen Malinowski
1
, Gregor Heinrich
2
, and Oliver Wendt
3
1
University of Frankfurt, Mertonstr. 17, 60325 Frankfurt/M., Germany
2
Fraunhofer IGD, Fraunhoferstr. 5, 64283 Darmstadt, Germany
3
University of Kaiserslautern, Postfach 3049, 67653 Kaiserslautern, Germany
Abstract. Recommendation systems are widely used on the Internet to assist
customers in finding the products or services that best fit their individual
preferences. While current implementations successfully reduce information
overload by generating personalized suggestions when searching for objects
such as books or movies, recommendation systems so far cannot be found in
another potential field of application: the personalized search for subjects such
as business partners or employees. This is astonishing as (1) the number of
CV-, assessment- and social network-data available on the Internet is growing
and (2) the complexity and scope of selecting the right partner is much higher
than when buying a book. We argue that recommendation systems
personalizing the search for people need to be grounded on two pillars: unary
attributes on the one hand and relational attributes on the other. We present a
framework meeting these requirements together with an outline of a first
prototypical implementation.
1 Introduction
Personalization systems such as recommender engines in recent years attracted the
interest of many researchers and practitioners. Since Resnick and Varian first
established the term “recommender system” in 1997 [26], researchers have been
improving recommendation quality and scalability of such systems by various means.
While some researchers merged content-based with collaborative filtering in order to
overcome sparsity problems and combine the advantages of both approaches [20]
[29], others focused on how to reduce the dimensionality of the user-item-matrix
underlying collaborative filtering approaches [30] [32]. Today, recommendation
systems successfully assist consumers on the Internet in finding products or objects
based on items similar to the ones the customer himself previously liked or based on
items that other customers similar to him liked in the past. However, personalization
systems are not yet applied when searching for people or subjects. Thus our research
question is: What are necessary theoretical enhancements for human recommender
systems? We argue that the various bilateral and relational aspects that need to be
considered when bringing individuals together imply extending existing approaches
Keim T., Malinowski J., Heinrich G. and Wendt O. (2005).
Personalizing the Search for Persons: A Recommender-based Approach.
In Proceedings of the 1st International Workshop on Web Personalisation, Recommender Systems and Intelligent User Interfaces, pages 125-134
DOI: 10.5220/0001421801250134
Copyright
c
SciTePress
by relational data. Building on existing theory and own prior research, we derive
concrete requirements and present an outline for a recommendation system
personalizing the search for individuals.
2 Research Motivation
Information technology in recent years has transformed (1) the ways people find work
as well as (2) the ways they effectively work together. With regard to the first aspect,
own longitudinal empirical research with the Top-1.000-companies in Germany as
well as with over 11.000 job seekers shows that the Internet has replaced print media
as the most important recruitment channel [16] [17]
1
. With 78% of all vacancies being
published within the career section of the corporate website and 49% of open jobs
being posted on Internet job-portals, IT-supported channels dominate print media
(30%) as a way to attract candidates. Also, over the years the ratio of actual hires
generated through job ads on the Internet rises reaching 58% in 2004 [17]. When
considering the later stages of the recruitment process such as the treatment of
incoming applications and the (pre-) selection of candidates, a diminished importance
of IS-support can be observed. However, as digital applications lower application
costs, the number of incoming (electronic) applications increases. Thus, companies
seek adapted IS-support for the selection stage in order to process the masses of
incoming applications efficiently.
85,0%
52,4%
37,4%
30,0%
16,6%
76,8%
48,9%
30,3%
23,1%
12,5%
0%
20%
40%
60%
80%
100%
Corporate Website Job-portal Print media Job Centers Others
2003 2004
52,8%
35,3%
4,3%
7,5%
58,2%
27,0%
5,2%
9,6%
0%
20%
40%
60%
80%
100%
Internet Print media Job Centers Others
2003 2004
Fig. 1. Ratio of vacancies published Fig. 2. Ratio of jobs filled
While this empirical research deals with how people find work, other research strands
are concerned with how information systems change the ways people effectively work
together once the candidate is recruited. Starting from Malone and Laubacher’s vision
of the “e-lance economy” [21], special attention was paid to the ways communication
channels and “discontinuities” of space, time and organizational boundaries
characteristic of virtual work influence collaboration patterns [2] [34]. Thus, as work
1
Companies selected based on revenues; between 151 and 196 companies responding between
2002 and 2004
126
in changing projects and organizational settings gains importance, individuals are
more frequently matched to new colleagues within their lifetime. Beside this, systems
for ad hoc short-term expert identification streamline the way knowledge is accessed
and exchanged between different projects or units beyond document management [1]
[9].
3 The Personalized Search for Persons
From these considerations that (1) matching situations within a person’s work history
will increase and (2) decision support for the matching of collaboration partners will
emerge, we started to develop a system for the personalized search for individuals. In
the following, we present requirements for such a person-recommender.
3.1 Requirements for Recommending Persons
Team configuration for work contexts has been analyzed by a variety of disciplines.
Typically, such problems are considered under the perspective of task-related and
social aspects [12], human and social capital or person-job fit and person-team or
person-organization fit [31]. Thus, successful team design needs to consider two
dimensions:
• The matching of individuals to tasks for which the candidate possesses the
skills and abilities to carry them out.
• The matching of individuals to other individuals with whom the person is
able to collaborate successfully.
This latter dimension has major implications for the design of a person-recommender
as we cannot consider the selection of candidates as a unilateral decision. While the
customer chooses the movie he wishes to watch and not vice versa, this is not the case
when recommending people. Selecting a candidate or partner is a bilateral selection
decision in which not only the attributes of the item or individual itself need to be
considered, but also the relationship between these items or individuals. In separation
to the former attributes that can be tied directly to the individual, that we refer to as
unary attributes, we denote the latter attributes as relational attributes. Thus, we retain
the following key differences when recommending subjects instead of objects:
• Recommending people is a bilateral process that needs to take into account
the preferences not only of a single person (the active user), but of several
persons.
• Even more, recommendations cannot be based on the attributes tied to the
items or persons in consideration only, but need to incorporate the
underlying relational structure by means of relational attributes.
• Finally, as every individual is considered to be unique, we cannot
recommend a single item or person several times such as in the case of a
127
movie or book. As every person can only be selected once, recommendations
on the item-level are not repeatable. Thus, recommendations cannot be
solely based on a user-item matrix but need to incorporate “content”-
elements such as the unary and relational attributes mentioned above.
3.2 Towards a Person-Recommender
On our way towards a person-recommender, we implemented two complementary
approaches: a CV-recommender and a social network browser that are both going to
be presented briefly hereunder. Afterwards we describe how both approaches can be
combined leading to a relational recommender system.
The CV-Recommender. In a first step, we built a system recommending CVs that
are similar to resumes previously selected by the same recruiter for a specific job-
profile considered. The probabilistic hybrid recommendation engine is based on a
latent aspect model that understands individual preferences as a convex combination
of preference factors [10] [11] [25]. As depicted in figure 3, the recruiter together
with the job description is represented in variable x, the preference factors being
modeled in variable z. In coherence with our prior considerations, the recruiter by
rating a candidate profile or CV with variable v = {"qualified", "not qualified"} does
not rate the person itself, but the sum of its attributes. These “content”-elements,
taken from the candidate's resume are composed of a quadruple such as
a=("mathematical skills", "diploma grade", "1.0", "University of Frankfurt"). Thus,
the rating value v depends indirectly on the position considered x and directly on the
candidate’s attributes a. With a set of observed values v for an attribute assessed by x
and assigned to a, we are able to estimate the model parameters using an Expectation
Maximization (EM) algorithm. A detailed description of the approach together with
validation results can be found in [4].
The Social Network Browser. As the CV-recommender is focused on what we
called unary attributes earlier, we modeled relational attributes in a complementary
approach. The network browser shown in figure 4 visualizes trusted social relations
that the user then can manually browse, filter the network and search for particular
nodes. The social relations used are recommendations between people based on
“historic” experience as well as swift trust assessments from candidate interviews via
video conferences. A more detailed description of the approaches to swift and historic
trust modeled within the system and their elicitation from a user community can be
found in [18] and [9]. When navigating the resulting network, by filtering and
searching techniques it is possible to identify relevant persons in the graph according
to different criteria. This way, important trusted actors in the network can be
identified either from an ego-centered position (of the searcher) or globally using
graph analysis methods such as shortest-path, relative importance and others also used
in social network analysis [33]. In addition, such filtering can initially apply relevant
competence criteria, which creates a trust network contextualized on the queried
competencies. The motivation of this idea is closely related to research on the
relationship between trust, interpersonal cooperation and organizational effectiveness
such as [3] or [14].
128
z
v
a
x
x Assessor and actual target attributex Assessor and actual target attribute
z Latent influencing factors of attribute
value
z Latent influencing factors of attribute
value
v Assessed target attribute valuev Assessed target attribute value
a Partner attribute a Partner attribute
Fig. 3. The probabilistic CV-Recommender Fig. 4. The network analyzer
4 Towards the Relational Person-Recommender
In order to meet the requirements previously defined, we need to combine the
predictive capabilities of the CV-recommender with the descriptive capabilities of the
network browser in an automatic setting. This is based on our previously defined
requirements where we stated that a person-recommender not only needs to consider
individual but also relational attributes. From a theoretical perspective, this is an
interesting idea as already Granovetter showed that labor market processes are rooted
in social relations [5]. Montgomery argued that the higher quality of information
gained from contact networks reduces frictions when entering a new job [22]. Also,
the reductions of attraction costs [28] and of screening costs have been mentioned as
advantages of partner identification over networks [19].
In order to build such a relational recommender, we developed a trust
computational model. Conforming to Richardson, Agrawal and Domingos (2003), we
assume that trust can be expressed in a singular value even though it is a complex and
multidimensional construct. (In the above network browser, we adopted this scheme
already by aggregating the values of the different trust dimension values.) Our current
research builds on trust propagation as demonstrated in [6]. Based on findings in the
literature and own theoretical considerations we defined three trust propagation and
prediction scenarios as depicted in Figure 5(a)-(c).
Figure 5(a) illustrates how the trust level between individuals A and C can be
inferred given the trust values t
AB
and t
BC
[27]. Figure 5(b) shows a typical
collaborative filtering approach to trust propagation that, based on three given
relations between four people, infers the missing trusted relation [6]. As a
complementary approach to trust propagation, we aim to directly combine individual
and relational attributes as depicted in Figure 5(c). Based on the existing individual
profiles A, B, C and D as well as a single existing trusted relationship t
A,B,
we will
calculate similarities between user pairs. Dependant on these distances d(x,y) as well
as the characteristics of the existing trusted relation t
A,B,
the system will recommend or
129
not a relationship between people unknown so far. We denote this approach as
similarity-based trust propagation. Our next steps include the further development of
our existing integrated prototype and its validation with real-life data. Also, we aim to
add social network data as an additional variable of the model and extend it by
different relation types.
A
B
C
t
AB
t
BC
t‘
AC
A B
C D
t
A
B
t
CD
t
A
D
t‘
CB
A B
C D
t
AB
t‘
CD
d(A,C) d(B,D)
a
A
a
B
a
C
a
D
Fig. 5(a). Direct trust
propagation
Fig. 5(b). Collaborative
trust propagation
Fig. 5(c). Similarity-based
trust propagation
As a basis for the predictive approach, we postulate two work hypotheses: The first
hypothesis is that the unary (i.e., propositional) and relational attribute structure
latently captures personal qualities that generate degrees of trust, possibly conditioned
on specific situations and roles. For instance, looking at a known relational
confidence attribute with a source A and a target B (e.g., A assesses B), it is predicted
that similar relations (with respect to type and weight) can be measured for sources
similar to A and targets similar to B. E.g., if A assesses B positively, C similar to A is
predicted to assess D similar to B positively, as well.
The second hypothesis of the approach is that some dimensions of trust are
transmissible through a referral network. This means, for example, that looking at
such a higher-order trust relation, A trusts B and B trusts C, again possibly
conditioned on a situation or role, trust from A to C can be predicted. This is the
conceptual basis of referral systems, such as ReferralWeb [15] [36]. The question is
what trust dimensions do exhibit this transitive behaviour to which degree.
In particular, the first hypothesis can be mapped to the emergent scientific area of
statistical relational learning (SRL), in which graph properties are learned from data
and the local graph topology surrounding newly observed nodes are predicted. In this
context, we note the work of Jensen, Neville and Wolfe [24] [35] [13] and of
Heckerman, Meek and Koller [8], as a basis for a generic social network prediction
algorithm. The second hypothesis is related to the friend-of-a-friend principle, which
is the basis for transitive trust relations and in fact the basis of the existing system
already.
Further, we plan to connect actors with documents to cluster actors by their
authorship and roles. This extends the idea of explicit profile creation to implicit
methods of profile creation, thus allowing for bootstrapping a real system by
connecting it to existing document bases. A scientific basis for work into this
direction can be found in [23]. Merging both the content and the social network into a
‘smart’ collaboration network to us seems a promising idea when considering the
many real-world knowledge management problems and applications. However,
several challenges appear when modelling profiles for the predictive approach to
partner matching. These are:
130
• the modelling of complementarity and compatibility for team building
scenarios. This includes the incorporation of research on matching different
personal traits with express expertise measures to optimize team staffing.
• the capturing of “inter-rater trust”. Within this functionality, the bias of a
rater will be used to remove bias from ratings and will also be incorporated
as a specific rater characteristic. This has been partly solved in our existing
Opal system via a matrix-based assessment browser as presented in [7].
• the resolution of disreputative scenarios. Situations in which candidates are
assessed badly must be resolved in a way that conserves overall integrity and
privacy in the community but that still allows marking negative experiences.
This is an often-encountered scenario where most rating-based systems
capitulate.
5 Validation approach
In order to validate our approach we currently design an experiment as part of a
student workshop. We plan to test the aspects of the described recommendation
framework in an incremental way. First, students are supposed to enter their CV data
into a web-based prototype. The data capturing hereby follows the same rules as it is
nowadays done in the various existing job-portals. The CV data together with
manually created ratings regarding the match of the students with several job-profiles
is then used as input to train the CV-recommender. Based on this training data the
system should then be able to predict the match between students and job profiles.
In a second step, Students will be asked to enter relational data into the prototype
such as their relations towards fellow students. The relations will be defined based on
its direction, duration and intensity. The captured data should then serve as input for
the trust computational model. Based on this training data the system should be able
to predict previously unknown relations. Finally we aim to combine the separate
results into an integrated approach for personalizing the search for persons.
6 Conclusion
In this paper we argued that recommendation systems so far personalize only the
search for objects, but not for subjects. We showed that theoretical extensions such as
the integration of relational as well as bilateral aspects into current approaches are
necessary in order to build a system personalizing the search for individuals. Based on
these requirements and building up on two implementations from previous research,
we presented an outline of a first existing prototype integrating both approaches into a
single system. Our next steps include the extension of this implementation as well as
its validation with real-life data as part of a student workshop to be carried out. The
objective is to enhance the matching quality of interpersonal partnership especially for
collaboration scenarios by building a bilateral as well as relational recommendation
engine personalizing the search for individuals.
131
Acknowledgements. We gratefully acknowledge the support of the European Union
under the Fifth Framework Programme Information Society Technologies (contract
number: IST-2000-28295).
References
1. Crowder, R., Hughes, G. and Hall, W. (2002) Approaches to Locating Expertise
Using Corporate Knowledge Int. J. Intell. Sys. Acc. Fin. Mgmt., 11 , 185-200.
2. DiTomaso, N. (2001) The loose coupling of jobs: the subcontracting of everyone,
in Berg, I. and Kalleberg, A.L. (Eds.) Sourcebook of Labor Markets: Evolving
Structures and Processes, Kluwer Academic/Plenum, New York, 247-270.
3. Dunphy, D. and Bryant, B. 1996 Teams: Panaceas or prescriptions for improved
performance? Human Relations, 49, 677-699.
4. Färber, F., Keim, T., and Weitzel, T. (2003) An Automated Recommendation
Approach to Personnel Selection, Proceedings of the 2003 Americas Conference
on Information Systems, Tampa.
5. Granovetter, M.S. (1985) Economic Action and Social Structure: the problem of
Embeddedness, American Journal of Sociology, Vol. 91, pp. 481 – 510.
6. Guha, R., Kumar, R., Raghavan, P. and Tomkins, A. (2004) Propagation of Trust
and Distrust, Proceedings of the WWW2004-Conference, May 17-22, New York,
USA, pp. 403-412.
7. Graham, M. and Kennedy, J. (2004) Exploring and Examining Assessment Data
via a Matrix Visualisation, Proceedings of the AVI 2004, Gallipoli, Lecce, Italy,
ACM Press. pp. 158-162.
8. Heckerman, D., Meek, C., and Koller, D. (2004) Probabilistic Models for
Relational Data, Technical Report MSR-TR-2004-30, Microsoft Research.
9. Heinrich, G. (2004) Teamarbeit nach Mass - Expertisemanagement in
Organisationsnetzwerken, Trendkompass Electronic Business - IT-Innovationen &
neue Prozesse im Unternehmenseinsatz, IRB-Verlag, Stuttgart.
10. Hofmann, T. (1999) Probabilistic latent semantic analysis, Proceedings of the 15th
Conference on Uncertainty in Artificial Intelligence (UAI), July 30-August 1,
Stockholm, Sweden, 289-296.
11. Hofmann, T. and Puzicha, J. (1999) Latent class models for collaborative filtering,
Proceedings of the 16th International Joint Conference on Artificial Intelligence,
July 31 – August 6, Stockhom, Sweden, 688-693.
12. Jackson, S.E. (1996) The consequences of diversity in multidisciplinary work
teams, in: West, M.A. Handbook of workgroup psychology, John Wiley & Sons,
Sussex.
13. Jensen, D., and Neville, J. (2002) Data Mining in Social Networks, NAS 2002.
14. Jones, G.R. and George, J.M. (1998) The experience and evolution of Trust:
Implications for Co-operation and Teamwork, The Academy of Management
Review, 23, 3, pp. 531-546.
132
15. Kautz, H., Selman, B. and Shah, M. (1997) Referral Web: Combining Social
Networks and Collaborative Filtering, Communications of the ACM, Vol. 40, no.
3, pp. 63-65.
16. Keim, T., König, W. and von Westarp, F. (2004) Bewerbungspraxis 2005 - Eine
empirische Untersuchung mit über 11.000 Stellensuchenden im Internet, Research
report, University of Frankfurt, 2004.
17. Keim, T., König, W., von Westarp, F. , Weitzel, T. and Wendt, O. “Recruiting
Trends 2005 - Eine empirische Untersuchung der Top-1000-Unternehmen in
Deutschland und von 1000 Unternehmen aus dem Mittelstand in Deutschland“,
Research report, University of Frankfurt, 2005.
18. Keim, T., Weitzel, T. (2005) An Integrated Framework for Online Partnership-
Building, Proceedings of the 38th Hawaiian International Conference on System
Sciences (HICSS-38), Hilton Waikoloa Village, Big Island, Hawaii.
19. Leicht, K. T. and Marx, J. (1997) The Consequences of Informal Job Finding for
Men and Women, Academy of Management Journal, 40, pp. 967-987.
20. Melville, P., Mooney, R.J. and Nagarajan, R. (2002) Content-boosted
collaborative filtering for improved recommendations, Proceedings of the 18th
National Conference on Artificial Intelligence, pp. 187-192.
21. Malone, T. W. and Laubacher, R. J. (1998) The Dawn of the E-Lance Economy,
Harvard Business Review, 76 (5), pp. 144-152
22. Montgomery, J. D. (1991) Social Networks and labor-market outcomes: toward an
economic analysis, American Economic Review, 81, pp. 1408-1417.
23. McCallum, A., Corrada-Emmanuel, A. and Wang, X. (2004) The Author-
Recipient-Topic Model for Topic and Role Discovery in Social Networks:
Experiments with Enron and Academic Email, Technical Report, UM-CS-2004-
096.
24. Neville, J., Adler, M. and Jensen, D. (2003) Clustering Relational Data Using
Attribute and Link Information, Text-Mining & Link-Analysis Workshop,
TextLink.
25. Popescul, A., Ungar, L.H., Pennock, D.M. and Lawrence, S. (2001) Probabilistic
models for unified collaborative and content-based recommendation in sparse-data
environments, Proceedings of the Seventeenth Conference on Uncertainty in
Artificial Intelligence, August 2-5, Seattle, USA, pp. 437-444.
26. Resnick, P. and Varian, H. R. (1997) Recommender systems. Communications of
the ACM, 40 (3), pp. 56-58.
27. Richardson, M., Agrawal, R. and Domingos, P. (2003) Trust management for the
semantic web, Proceedings of the Second International Semantic Web
Conference, October 20-23, Sanibel Islands, USA, pp. 351-368.
28. Russo, G., Rietveld, P., Nijkamp, P. and Gorter, C. (2000) Recruitment channel
use and applicant arrival: An empirical analysis, Empirical economics, 25, pp.
673-697.
29. Sarwar, B., Karypis, G., Konstan, J. and Riedl, J. (2000) Analysis of
recommendation algorithms for e-commerce, Proceedings of the ACM
Conference on Electronic Commerce, pp. 158-167.
30. Sarwar, B., Karypsis, G., Konstan, J. A. and Riedl, J. (2000) Application of
dimensionality reduction in recommender system – A case study, Proceedings of
the ACM WebKDD 2000 Web Mining for E-Commerce Workshop, ACM, New
York.
133
31. Schneider, B., Kristof-Brown, A., Goldstein, H. W. and Smith, D. B. (1997) What
is this thing called fit?, in Anderson, N. and Herriot, P. (Eds.): International
handbook of Selection and Assessment, John Wiley & Sons, pp. 393-412.
32. Ungar, L. and Foster, D. (1998) Clustering methods for collaborative filtering,
Proceedings of the Workshop on Recommendation Systems, AAAI Press, Menlo
Park, California.
33. Wasserman, S. and Faust, K. (1994) Social Network Analysis: Methods and
Applications, Cambridge University Press.
34. Watson-Manheim, M.B., Crowsten, K. and Chudoba, K.M. (2002) Discontinuities
and continuities: a new way to understand virtual work, Information Technology
& People, Vol. 15(3), pp. 191-209.
35. Wolfe, A. and Jensen, D. (2004) Playing Multiple Roles, Discovering Overlapping
Roles in Social Networks, ICML 2004 Workshop on Statistical Relational
Learning and its Connections to Other Fields.
36. Yolum, P. and Singh, M. P. (2003) Emergent properties of referral systems,
Proceedings of the 2nd International Joint Conference on Autonomous Agents and
MultiAgent Systems (AAMAS), ACM Press.
134
