presented a per-query algorithmic approach that leverages fundamental
retrieval principles such as pseudo-feedback-based relevance modeling. 24
That is, the weak ranker observes the
documents most highly ranked by the
strong ranker, uses them to induce a
model of relevance for the query, and
uses this model for ranking. We demonstrated the merits of our approach
in improving the search effectiveness
of the weak ranker using TREC datae—
a suite of datasets used for evaluation
of retrieval methods.
Strategic Users
The previous sections emphasized
the need to account for the incentives of strategic parties when devising major data science applications
such as search (ranking), recommendation, and prediction (specifically,
regression). The strategic parties
were associated with stakeholders
such as content providers or prediction experts. The users of the systems
were not strategic; they were the ones
posting a query, asking for recommendation, or interested in prediction. In a sense, the users were the
products the strategic parties were
aiming to attract or to serve better
than their competitors. However,
users may have their own incentives
that might cause them to potentially
not follow recommendations by the
system or to provide dishonest inputs to the system. Next, we briefly
describe one of our recent works that
addresses such aspects.
Incentive-compatible explore and
exploit. Recommendation systems
based on the interleaving of explora-
tion and exploitation paradigm have
a two-way relationship with their
customers—on the one hand they
provide recommendations while on
the other hand they use customers as
their source of information. This re-
ality leads to an important challenge:
a user might not accept a recommen-
dation if he/she believes the recom-
mendation is done to benefit explo-
ration rather than be the optimal
one given current information. The
challenge is to devise a system that
will behave close to optimal, but will
be also incentive compatible, that is,
e https://trec.nist.gov/
We have recently addressed this
challenge in designing recommenda-
tion systems, specifically for social
networks, where users can (partly)
observe each other. 1,f In particular, we
investigated the conditions on the so-
cial network that allow for asymptoti-
cally optimal outcomes. Our results
show that for reasonable networks,
where each user can observe many of
the other users, but where still most
users cannot see most of the other
users, an incentive compatible and
approximately optimal recommenda-
tion system does exist.
The literature on incentivizing exploration relevant to our study can be
viewed as an extension of the celebrated multi-armed bandit problems8 to
deal with settings where exploration
can only be done by self motivated myopic agents and a central planner must
incentivize exploration. The tension between the objective of the mediator (the
recommendation engine) and the individual agents in a multi-armed bandit
context was first introduced in Kremer
et al., 22 who study this in a very simple
setting. They identify an incentive compatible scheme with which a central
mediator with commitment power can
asymptotically steer the users toward
taking the optimal action. This exciting
news has been extended in Mansour et
al. 26, 27 to several more elaborate bandit
settings and to additional optimization
criteria such as regret minimization.
Whereas both of these papers account
for agents’ incentives and in particular the misalignment of incentives of
the agents and the mediator they ignore other societal aspects. In particular, these papers make an implicit assumption that agents cannot see nor
communicate with any other agent.
Needless to say, the assumption that
agents have no knowledge of each other, and cannot see the actions chosen
by their neighbors, is unrealistic. Che
and Horner9 also study mechanisms
for social learning through a mediator.
Similar to Kremer et al., 22 they assume
agents are short lived and do not observe each other and the mechanism
f The pioneering work on this challenge
assumed no user communication. 9. 22
A recommender
system should be
able to deal in
a fair way with
strategic content
providers,
each of which
aims at maximizing
its exposure.