In order to clarify the nature of profiling technologies, some crucial distinctions have to be made between different types of profiling practices, apart from the distinction between the construction and the application of profiles. The main distinctions are those between bottom-up and top-down profiling (or supervised and unsupervised learning), and between individual and group profiles.
Supervised and unsupervised learning Profiles can be classified according to the way they have been generated . On the one hand, profiles can be generated by testing a hypothesized correlation. This is called top-down profiling or
supervised learning. This is similar to the methodology of traditional scientific research in that it starts with a hypothesis and consists of testing its validity. The result of this type of profiling is the verification or refutation of the hypothesis. One could also speak of deductive profiling. On the other hand, profiles can be generated by exploring a data base, using the
data mining process to detect patterns in the data base that were not previously hypothesized. In a way, this is a matter of generating hypothesis: finding correlations one did not expect or even think of. Once the patterns have been mined, they will enter the loop – described above – and will be tested with the use of new data. This is called
unsupervised learning. Two things are important with regard to this distinction. First, unsupervised learning algorithms seem to allow the construction of a new type of knowledge, not based on hypothesis developed by a researcher and not based on causal or motivational relations but exclusively based on stochastical correlations. Second, unsupervised learning algorithms thus seem to allow for an inductive type of knowledge construction that does not require theoretical justification or causal explanation . Some authors claim that if the application of profiles based on computerized stochastical pattern recognition 'works', i.e. allows for reliable predictions of future behaviours, the theoretical or causal explanation of these patterns does not matter anymore . However, the idea that 'blind' algorithms provide reliable information does not imply that the information is neutral. In the process of collecting and aggregating data into a database (the first three steps of the process of profile construction), translations are made from real-life events to
machine-readable data. These data are then prepared and cleansed to allow for initial computability. Potential bias will have to be located at these points, as well as in the choice of algorithms that are developed. It is not possible to mine a database for all possible linear and non-linear correlations, meaning that the mathematical techniques developed to search for patterns will be determinate of the patterns that can be found. In the case of machine profiling, potential bias is not informed by common sense prejudice or what psychologists call stereotyping, but by the computer techniques employed in the initial steps of the process. These techniques are mostly invisible for those to whom profiles are applied (because their data match the relevant group profiles).
Individual and group profiles Profiles must also be classified according to the kind of subject they refer to. This subject can either be an individual or a group of people. When a profile is constructed with the data of a single person, this is called individual profiling . This kind of profiling is used to discover the particular characteristics of a certain individual, to enable unique identification or the provision of personalized services. However, personalized servicing is most often also based on group profiling, which allows categorisation of a person as a certain type of person, based on the fact that her profile matches with a profile that has been constructed on the basis of massive amounts of data about massive numbers of other people. A group profile can refer to the result of data mining in data sets that refer to an existing community that considers itself as such, like a religious group, a tennis club, a university, a political party etc. In that case it can describe previously unknown patterns of behaviour or other characteristics of such a group (community). A group profile can also refer to a category of people that do not form a community, but are found to share previously unknown patterns of behaviour or other characteristics . In that case the group profile describes specific behaviours or other characteristics of a category of people, like for instance women with blue eyes and red hair, or adults with relatively short arms and legs. These categories may be found to correlate with health risks, earning capacity, mortality rates, credit risks, etc. If an individual profile is applied to the individual that it was mined from, then that is direct individual profiling. If a group profile is applied to an individual whose data match the profile, then that is indirect individual profiling, because the profile was generated using data of other people. Similarly, if a group profile is applied to the group that it was mined from, then that is direct group profiling . However, in as far as the application of a group profile to a group implies the application of the group profile to individual members of the group, it makes sense to speak of indirect group profiling, especially if the group profile is non-distributive.
Distributive and non-distributive profiling Group profiles can also be divided in terms of their distributive character . A group profile is distributive when its properties apply equally to all the members of its group: all bachelors are unmarried, or all persons with a specific gene have 80% chance to contract a specific disease. A profile is non-distributive when the profile does not necessarily apply to all the members of the group: the group of persons with a specific postal code have an average earning capacity of XX, or the category of persons with blue eyes has an average chance of 37% to contract a specific disease. Note that in this case the chance of an individual to have a particular earning capacity or to contract the specific disease will depend on other factors, e.g. sex, age, background of parents, previous health, education. It should be obvious that, apart from tautological profiles like that of bachelors, most group profiles generated by means of computer techniques are non-distributive. This has far-reaching implications for the accuracy of indirect individual profiling based on data matching with non-distributive group profiles. Quite apart from the fact that the application of accurate profiles may be unfair or cause undue stigmatisation, most group profiles will not be accurate. == Applications ==