leaves the active user in charge of driving the system in a way that contextual
requirements are met. Proactive design
follows the logic of prescriptive analytics where the system suggests actions
or even takes them on behalf of the
active user. Whereas the proactive design introduces higher requirements
for data availability and quality, it also
provides a stronger basis for pursuing
serendipity and diversity.
Multi-dimensional analytics. As noted,
contextuality and the goal of supporting
systemic perspective call for multidimensional analysis with several parallel relevance algorithms and various
data about the actors. Extant research
on recommender systems shows that
the perceived relevance of recommendations and the perceived usefulness
of the recommender system increase
when multiple different recommendation strategies are in parallel play. For
example, Hupa et al. 18 discuss the advantages of multidimensional social-network recommenders to increase
interdisciplinary collaboration. Tsai
and Brusilovsky37 used four different
recommender engines to support identifying new academic collaborators using conference data (topic similarity,
social similarity, interest similarity, and
geographical distance) and showed that
the users who are able and willing to
use multiple engines in parallel receive
more relevant recommendations.
Although the triangulation of these
approaches can already help identification of relevant matches, we call
for consideration of additional perspectives in the prescriptive analytics for people recommendations. For
example, geographical distance must
be considered, especially in long-term
collaboration. Level of seniority (or expertise) can affect the preferred symmetry of benefit and trust: matching
for mentoring or advisory relationships calls for high difference in expertise, while matching for a production team often demands relatively
equal levels of seniority. The personalities and organization cultures need
to be similar enough to enable matches that are sustainable in the long
term (for example, no conflicts due to
drastically different ways of working
or level of commitment).
Overall, developing analytics pro-
cedures and recommender engines
for different perspectives is a step to-
ward so-called hybrid recommender
systems. 37 At the same time, this intro-
duces practical challenges. In addition
to the axiomatic data availability issue,
defining the logic in which the differ-
ent analytical functions are combined
in a context-sensitive manner requires
deliberate research. Because PSM
can potentially be affected by such a
broad range of human features, all of
them cannot practically be embedded
in the algorithm design. We can only
call for multidisciplinary research col-
laboration where social scientific un-
derstanding would support the iden-
tification of top-priority factors and
formalizing this vast space.
User-system cooperation for decision
support. Due to the dynamic and inherently complex nature of PSM, we argue
that the matching decisions cannot
merely be automated or offloaded to algorithms. For example, machine learning generally produces relevant results
only if initialized with good-quality
training data and well-defined goals,
whereas human reasoning is suited for
multi-faceted challenges where the desired pattern is unknown a priori. The
different limitations and strengths in
the human and computational analytical capabilities call for effective
collaboration between computational
intelligence (deep yet narrow) and human intelligence (broad yet shallow).
This relates to the general notions of
augmented intelligence and human-in-the-loop thinking. As this approach
has already shown its power in, for
example, classification problems, the
complexity of PSM offers an even more
opportune application area.
The human-in-the-loop approach
has been envisionedf as useful, for ex-
ample, when: (a) the cases that need to
be identified are rare (class imbalance,
for example, rare type of collaboration);
(b) the cost of error is high (for exam-
ple, time spent on browsing irrelevant
matching options); (c) human anno-
tations are already used (for example,
recruiting processes); and (d) generic
pre-trained models exist but need to be
customized. The holistic thinking and
contextual adaptability of the user are
needed, for example, to steer the deep
yet narrow algorithms (for example,
f http://bit.ly/2mvBU5m
refinement of what an ideal match
is, or prioritizing the sought features
for each matching case) and to make
sense of and choose between the re-
sulting recommendations. Particularly
when considering non-experts as us-
ers of recommender systems, we need
user interface solutions that support
decision-making with alternative op-
tions, a multi-dimensional systemic
viewpoint, and ways to communicate
and deal with the algorithmic uncer-
tainty that this application area entails.
Figure 5 outlines the potential col-
laboration points along the computa-
tional analytics process. First, we need
methods that guide the user in selecting
appropriate training data and analyti-
cal goals to enable accurate profiling
and, eventually, predictions. Semi-su-
pervised approaches could allow train-
ing with a very small amount of labeled
training data, minimizing both the cold
start problem and need for manual work
by the user. Second, we need feedback
from the user regarding which factors
are of top priority in the current match-
ing case. Third, we need to support user
exploration for enacted sensemaking,
which has been found to be important
both in visual analytics in general6 and
visual network analytics in particular. 17
Subscribing to the call for transpar-
ency in AI and algorithmic systems, we
argue that also PSMs should be able to
better explain the reasoning behind the
recommendations. For example, the
uncertainty of the prediction should
be translated into human-comprehen-
sible forms so that the user can trust
them, is able to assess the factual rele-
vance of the recommendation, and can
adjust their preferences accordingly.
Throughout the process, the cost
in terms of burdening the user must
be in line with the benefits of using
the system. This raises the question of
what user input and feedback is sufficient to hone the algorithms while
causing minimal burden with tedious
tasks for the user. In other words, it
introduces a trade-off between the
certainty of the matching decision
and invested user effort.
An example of interactive visualization that can support exploration of
relevant other people is the Conference Navigator. 38 The recommended
conference talks are presented as sets,
each of which are identified by a recom-