Bias in algorithmic filtering and personalization 213

factor in the control of society. Giving citizens a choice in second phase focuses on the construction of a user profile on
ideas and information is as important as giving them choice basis of the user data. Here the collected and stored data are
in politics. analyzed and processed. In the final phase, the compiled user
In 2005, the Pew Internet and American Life Project profile is used in the actual web service, for instance a cus-
reported on the rise of search engines, and surveyed users’ tomized newsfeed in a social networking site, personalized
knowledge of how they worked. It concluded that ‘‘search results in a search engine query, or recommended products in
engines are attaining the status of other institutions—legal, an e-commerce site.
medical, educational, governmental, journalistic—whose A system can build a user profile in two ways:
performance the public judges by unusually high standards,
• Explicitly: the user customizes the information source
because the public is unusually reliant on them for prin-
himself. The user can register his interests or demo-
cipled performance’’ (Fallows 2005). Personalization and
graphic information before the personalization starts.
other forms of algorithmic filtering are thus ‘‘replacing the
The user can also rate topics of interest.
traditional repositories that individuals and organizations
• Implicitly: the system determines what the user is
turn to for the information needed to solve problems and
interested in through various factors, including web
make decisions’’ (Mowshowitz and Kawaguchi 2002). The
usage mining (i.e., previous interaction with the system
services that employ such algorithms are gateways that act
such as clickthroughs, browsing history, previous
as intermediaries between information sources and infor-
queries, time spend reading information about a
mation seekers. They play a vital role in how people plan
product), IP address, cookies, session id’s, etc.
and live their lives. Since access to information is a value,
and online filters allow or block access to information, Explicit user information collection will allow the user
building these algorithms is not only a technical matter, but to know that the personalization is taking place and he can
a political one as well. Before discussing how bias can tailor it to his needs. However, one problem with explicit
manifest itself in personalization, it is important to first feedback is that it places an additional burden on the user.
understand how personalization works. Because of this, or because of privacy concerns, the user
may not choose to participate. It is also known that users
may not accurately report their own interests or demo-
Personalization: a technical overview graphic data, or, since the profile remains static whereas the
user’s interests may change over time (Gauch et al. 2007).
Most personalization systems are based on some type of user Implicit user information collection, on the other hand,
profile, a data instance of a user model that is applied to does not require any additional intervention by the user
adaptive interactive systems. User profiles may include during the process of constructing profiles. It also auto-
demographic information, (e.g., name, age, country, educa- matically updates as the user interacts with the system. One
tion level), and may also represent the interests or prefer- drawback of implicit feedback techniques is that they can
ences of either a group of users or a single person. In general, typically only capture positive feedback. When a user
the goal of user profiling is to collect information about the clicks on an item or views a page, it seems reasonable to
subjects in which a user is interested, and the length of time assume that this indicates some user interest in the item.
over which they have exhibited this interest, in order to However, it is not clear that when a user fails to examine
improve the quality of information access and infer the user’s some data item it is an indication of disinterest (Gauch
intentions. As shown in Fig. 2, the user profiling process et al. 2007).
generally consists of three main phases. First, an information Different techniques can be used to make suggestions to
collection process is used to gather raw information about the users on which information is relevant for them. Recom-
user. Depending on the information collection process mendation systems try to analyze how a user values certain
selected, different types of user data can be extracted. The products or services and then predict what the user will be
interested in next. A recommendation mechanism typically
Explicit does not use an explicit query but rather analyses the user
Info
Data Profile Technology or
context (e.g., what the user has recently purchased or read,
Collection Constructor application and, if available, a user profile (e.g., the user likes mystery
Implicit
Info novels). Then the recommendation mechanism presents to
the user one or more descriptions of objects (e.g., books,
Personalized people, movies) that may be of interest (Adomavicius et al.
User services
2005; Garcia-Molina et al. 2011).
Fig. 2 User profile construction for personalization (adapted from If this recommendation is done solely by analyzing the
Gauch et al. 2007) associations between the user’s past choices and the

123