Decaying Relevance of Clinical Data when Predicting Future Decisions

Muthuraman Alagappan, Mary K. Goldstein, Steven M. Asch, Russ B. Altman, Jonathan H. Chen

Abstract: Classical learning methods are often based on a simple but restrictive assumption: The present and future data are generated according to the same distributions. This hypothesis is particularly convenient when it comes to developing theoretical guarantees that the learning is accurate. However, it is not realistic from the point of view of applicative domains that have emerged in the last years.In this thesis, we focus on four distinct problems in artificial intelligence, that have mainly one common point: All of them imply knowledge transfer from one domain to the other. The first problem is analogical reasoning and concerns statements of the form "A is to B as C is to D". The second one is transfer learning and involves classification problem in situations where the training data and test data do not have the same distribution (nor even belong to the same space). The third one is data stream mining, ie. managing data that arrive one by one in a continuous and high-frequency stream with changes in the distributions. The last one is collaborative clustering and focuses on exchange of information between clustering algorithms to improve the quality of their predictions.The main contribution of this thesis is to present a general framework to deal with these transfer problems. This framework is based on the notion of Kolmogorov complexity, which measures the inner information of an object. This tool is particularly adapted to the problem of transfer, since it does not rely on probability distributions while being able to model the changes in the distributions.Apart from this modeling effort, we propose, in this thesis, various discussions on aspects and applications of the different problems of interest. These discussions all concern the possibility of transfer in multiple domains and are not based on complexity only.