## AbstractStatistical inference forms the backbone of modern science. It is often viewed as giving an objective validation for hypotheses or models. Perhaps for this reason the theory of statistical inference is often derived with the assumption that the "truth" is within the model family. However, in many real-world applications the applied statistical models are incorrect. A more appropriate probabilistic model may be computationally too complex, or the problem to be modelled may be so new that there is little prior information to be incorporated. However, in statistical theory the theoretical and practical implications of the incorrectness of the model family are to a large extent unexplored. This thesis focusses on conditional statistical inference, that is, modeling of classes of future observations given observed data, under the assumption that the model is incorrect. Conditional inference or prediction is one of the main application areas of statistical models which is still lacking a conclusive theoretical justification of Bayesian inference. The main result of the thesis is an axiomatic derivation where, given an incorrect model and assuming that the utility is conditional likelihood, a discriminative posterior yields a distribution on model parameters which best agrees with the utility. The devised discriminative posterior outperforms the classical Bayesian joint likelihood-based approach in conditional inference. Additionally, a theoretically justified expectation maximization-type algorithm is presented for obtaining conditional maximum likelihood point estimates for conditional inference tasks. The convergence of the algorithm is shown to be more stable than in earlier partly heuristic variants. The practical application field of the thesis is inference of relevance from eye movement signals in an information retrieval setup. It is shown that relevance can be predicted to some extent, and that this information can be exploited in a new kind of task, proactive information retrieval. Besides making it possible to design new kinds of engineering applications, statistical modeling of eye tracking data can also be applied in basic psychological research to make hypotheses of cognitive processes affecting eye movements, which is the second application area of the thesis.
[Edit] |